What is the Agent Checkout Rate?
The Agent Checkout Rate is the share of AI-agent visits to a store that reach checkout. Checkout, not purchase, because payment is often handed back to the agent's person and often happens where the store's own scripts cannot see it. Read beside a randomly held-back control group, it compares agents a system acted for with agents it left alone.
Last updated
Why measure checkout and not the purchase?
Because checkout is the last step a store can reliably see an agent take. Payment is often handed back to the person. And payment often happens on a page the store's own scripts do not run on.
Two things push the visible end of the visit earlier. The first is the mandate. In the agent bench, agents told to report back to their person never reached the cart; agents allowed to buy placed orders. The second is where payment lives. On a Shopify store, Cromanion's tag does not load on the checkout, which Shopify hosts, so the last page it sees is the cart. A payment button inside an iframe, such as PayPal's, is not seen either.
Checkout is therefore the conversion an agent visit can be counted on, on every store. When a real purchase is observed as well, it counts too.
What counts as reaching checkout?
A visit counts when either of two things happened. It reached checkout: it viewed a checkout page, or clicked toward one. Or a conversion was observed: a goal the store validated fired, or the server confirmed a purchase.
It reached checkout
A pageview of the visit was a checkout page. Or a click went toward one: the cart's checkout button, Shopify's "Buy it now" and express buttons, or a link whose destination is a checkout page. A click counts even if the checkout then refused the visitor, because the agent took the step.
A conversion was observed
A goal the store defined and validated fired, or the store's server confirmed a purchase. Observed conversions are certainty, not estimates, and they count whatever pages the visit saw.
Which visits count as agent visits?
Every visit that was not classified as a person. That covers agents that prove who they are, with a verified Web Bot Auth signature or one of Cloudflare's verified AI categories (assistant, search or crawler), and browsers recognized as driven by software from how they behave. Visits by the store's own team are excluded.
Many agents are not declared. Muse, for example, ran an ordinary browser and was recognized from behavior: a pointer that only ever jumps. An agent that is not recognized is counted as a person, and is missing from this rate.
How does a control group make the rate a comparison?
Cromanion shows two groups side by side. Agent visits where the system acted, and agent visits in the permanent holdout, where it held back at the same moment. Visitors are assigned at random, so on average the two groups differ only by whether the system acted.
Each group is shown as a count: X of Y visits reached checkout. The window is the visits that started in the last 24 hours.
A visit enters the comparison only when it reaches a moment where the system could act. An agent visit that never does is in neither group.
The acting group counts every visit where the system decided to act, whether or not anything was displayed. The control group can only be "would have acted", so this is the fair match for it. It measures the decision, not what reached a screen.
Why is the Agent Checkout Rate never shown as a lift?
Because the counts are small and a lift is a causal claim. Below 30 agent visits in each group, only raw counts are shown. Above it, each group's rate appears in words beside its counts. At no size does the comparison claim an effect.
Small groups produce large, meaningless gaps. A rate from 7 visits against a rate from 3 says nothing about either.
Before-and-after comparisons are worse. A handful of lab runs on Cromanion's own test store looked better after a change: four runs on each side. They are anecdotal. There was no concurrent control, and the agents keep memory across conversations, so the second runs were not fresh visitors. That is exactly the comparison a holdout exists to replace.
What can the Agent Checkout Rate not see?
Anything that happens where the tag does not run, and any agent it did not recognize. A payment button in an iframe is missed. An agent told to report back may recommend the store without reaching checkout. And one agent visit can occasionally split into two sessions.
Payment in an iframe
A button inside an embedded payment frame, such as PayPal's, is invisible to the page's scripts. A visit that paid there counts only if a conversion was also observed.
Report-back agents
An agent asked to recommend rather than buy may stop before the cart. Its visit counts as not reaching checkout, even when its person later buys.
Unrecognized agents
An agent the detection missed is counted as a person. The rate covers recognized agents only.
Split visits
Muse visits observed on 2026-09-25 sometimes landed in two sessions. A split visit can count twice, once without the checkout.
Common questions
Is the Agent Checkout Rate the same as a conversion rate?
No. A conversion rate counts purchases or goals. The Agent Checkout Rate counts the last step an agent can be seen taking, which for an agent is often checkout, and adds any conversion that was observed. For agents, it is the closest visible stand-in for a sale.
Does a store in learn mode have an Agent Checkout Rate?
It has half of one. A store that is only learning never acts, so its acting group stays at zero. Its control group still fills, which shows how agents do on the store with nothing changed.
See it on your own site
Paste one tag. Cromanion crawls your site, watches real sessions in Learn mode, and only acts when you switch it on — with a permanent 10% holdout proving what it caused. Free to start, no credit card: a 14-day or 1,000-session live trial, then keep measuring for free.